Do not emit __gnu_lto_v1 symbol.

2019-07-22  Martin Liska  <mliska@suse.cz>

* config/avr/avr.c (avr_asm_output_aligned_decl_common): Update
comment.
* toplev.c (compile_file): Do not emit __gnu_lto_v1 symbol.
2019-07-22  Martin Liska  <mliska@suse.cz>

* config/pa/stublib.c: Remove stub symbol __gnu_lto_v1.
* config/pa/t-stublib: Likewise.
2019-07-22  Martin Liska  <mliska@suse.cz>

* simple-object-elf.c (simple_object_elf_copy_lto_debug_sections):
Do not search for gnu_lto_v1, but search for first '\0'.

From-SVN: r273662

Simplify LTO section format.

2019-07-22  Martin Liska  <mliska@suse.cz>

* lto-section-in.c (lto_get_section_data):
Use new function get_compression.
* lto-streamer-out.c (produce_lto_section): Use
set_compression to encode compression algorithm.
* lto-streamer.h (struct lto_section): Do not
use bitfields in the format.

From-SVN: r273661

Make a warning for -Werror=wrong-language (PR driver/91172).

2019-07-22  Martin Liska  <mliska@suse.cz>

PR driver/91172
* opts-common.c (decode_cmdline_option): Decode
argument of -Werror and check it for a wrong language.
* opts-global.c (complain_wrong_lang): Remove such case.
2019-07-22  Martin Liska  <mliska@suse.cz>

PR driver/91172
* gcc.dg/pr91172.c: New test.

From-SVN: r273660

[ARC] Fix emitting TLS symbols.

When storing a TLS symbol to memory, always use an intermediate register to load it.

gcc/
xxxx-xx-xx  Claudiu Zissulescu  <claziss@synopsys.com>

* config/arc/arc.c (prepare_move_operands): Always use an
intermediate register when storing a TLS symbols.

gcc/
xxxx-xx-xx  Claudiu Zissulescu  <claziss@synopsys.com>

* gcc/testsuite/gcc.target/arc/tls-2.c: New test.
* gcc/testsuite/gcc.target/arc/tls-3.c: Likewise.

From-SVN: r273657

Daily bump.

From-SVN: r273656

re PR c++/67853 (decltype of parenthesized xvalue does not correctly yield rvalue-reference)

PR c++/67853
* g++.dg/cpp0x/decltype72.C: New test.

From-SVN: r273652

or1k: only force reg for immediates

The force_reg in or1k_expand_compare is hard coded for SImode, which is fine as
this used to only be used on SI expands.  However, with FP support this will
cause issues.  In general we should only force the right hand operand to a
register if its an immediate.  This patch adds an condition to check for that.

gcc/ChangeLog:

* config/or1k/or1k.c (or1k_expand_compare): Check for int before
force_reg.

From-SVN: r273651

or1k: Initial support for FPU

This adds support for OpenRISC hardware floating point instructions.
This is enabled with the -mhard-float option.

Double-prevision floating point operations work using register pairing as
specified in: https://openrisc.io/proposals/orfpx64a32.  This has just been
added in the OpenRISC architecture specification 1.3.
This is enabled with the -mdouble-float option.

Not all architectures support unordered comparisons so an option,
-munordered-float is added.

Currently OpenRISC does not support sf/df or df/sf conversions, but this has
also just been added in architecture specification 1.3.

gcc/ChangeLog:

* config.gcc (or1k*-*-*): Add mhard-float, mdouble-float, msoft-float
and munordered-float validations.
* config/or1k/constraints.md (d): New register constraint.
* config/or1k/predicates.md (fp_comparison_operator): New.
* config/or1k/or1k.c (or1k_print_operand): Add support for printing 'd'
operands.
(or1k_expand_compare): Normalize unordered comparisons.
* config/or1k/or1k.h (reg_class): Define DOUBLE_REGS.
(REG_CLASS_NAMES): Add "DOUBLE_REGS".
(REG_CLASS_CONTENTS): Add contents for DOUBLE_REGS.
* config/or1k/or1k.md (type): Add fpu.
(fpu): New instruction reservation.
(F, f, fr, fi, FI, FOP, fop): New.
(<fop><F:mode>3): New ALU instruction definition.
(float<fi><F:mode>2): New conversion instruction definition.
(fix_trunc<F:mode><fi>2): New conversion instruction definition.
(fpcmpcc): New code iterator.
(*sf_fp_insn): New instruction definition.
(cstore<F:mode>4): New expand definition.
(cbranch<F:mode>4): New expand definition.
* config/or1k/or1k.opt (msoft-float, mhard-float, mdouble-float,
munordered-float): New options.
* doc/invoke.texi: Document msoft-float, mhard-float, mdouble-float and
munordered-float.

From-SVN: r273650

or1k: Add mrori option, fix option docs

gcc/ChangeLog:

* config.gcc (or1k*-*-*): Add mrori and mror to validation.
* doc/invoke.texi (OpenRISC Options): Add mrori option, rewrite all
documenation to be more clear.
* config/or1k/elf.opt (mboard=, mnewlib): Rewrite documentation to be
more clear.
* config/or1k/or1k.opt (mrori): New option.
(mhard-div, msoft-div, mhard-mul, msoft-mul, mcmov, mror, msext,
msfimm, mshftimm): Rewrite documentation to be more clear.
* config/or1k/or1k.md (insn_support): Add ror and rori.
(enabled): Add conditions for ror and rori.
(rotrsi3): Replace condition for shftimm with ror and rori.

gcc/testsuite/ChangeLog:

* gcc.target/or1k/ror-4.c: New file.
* gcc.target/or1k/shftimm-1.c: Update test from rotate to shift
as the shftimm option no longer controls rotate.

From-SVN: r273649

or1k: Fix issues with msoft-div

Fixes bad assembly logic with software divide as reported by Richard Selvaggi.
Also, add a basic test to verify the soft math works when enabled.

gcc/testsuite/ChangeLog:

PR target/90362
* gcc.target/or1k/div-mul-3.c: New test.

libgcc/ChangeLog:

PR target/90362
* config/or1k/lib1funcs.S (__udivsi3): Change l.sfeqi
to l.sfeq and l.sfltsi to l.sflts equivalents as the immediate
instructions are not available on every processor.  Change a
l.bnf to l.bf to fix logic issue.

From-SVN: r273648

or1k: Fix code quality for volatile memory loads

Volatile memory does not match the memory_operand predicate.  This
causes extra extend/mask instructions instructions when reading
from volatile memory.  On OpenRISC loading volatile memory can be
treated the same as regular memory loads which supports combined
sign/zero extends.  Fixing this eliminates the need for extra
extend/mask instructions.

This also adds a test provided by Richard Selvaggi which uncovered the
issue while we were looking into another issue.

gcc/ChangeLog:

PR target/90363
* config/or1k/or1k.md (zero_extend<mode>si2): Update predicate.
(extend<mode>si2): Update predicate.
* gcc/config/or1k/predicates.md (volatile_mem_operand): New.
(reg_or_mem_operand): New.

gcc/testsuite/ChangeLog:

PR target/90363
* gcc.target/or1k/swap-1.c: New test.
* gcc.target/or1k/swap-2.c: New test.

From-SVN: r273647

[PPC] Fix bootstrap for non-SVR4 targets.

The recent change to move code into the new rs6000-call.c file is missing a
default value for the TARGET_NO_PROTOTYPE value (which only affects targets
that don’t include svr4.h).  Fixed by moving the fallback setting from
rs6000.c (which has no uses now) to rs6000-call.c.

2019-07-21  Iain Sandoe  <iain@sandoe.co.uk>

* config/rs6000/rs6000.c (TARGET_NO_PROTOTYPE): Move from here...
* config/rs6000/rs6000-call.c: ... to here.

From-SVN: r273646

re PR libfortran/91030 (Poor performance of I/O -fconvert=big-endian)

2019-07-21  Thomas König  <tkoenig@gcc.gnu.org>

PR libfortran/91030
* gfortran.texi (GFORTRAN_FORMATTED_BUFFER_SIZE): Document
(GFORTRAN_UNFORMATTED_BUFFER_SIZE): Likewise.

2019-07-21  Thomas König  <tkoenig@gcc.gnu.org>

PR libfortran/91030
* io/unix.c (BUFFER_SIZE): Delete.
(BUFFER_FORMATTED_SIZE_DEFAULT): New variable.
(BUFFER_UNFORMATTED_SIZE_DEFAULT): New variable.
(unix_stream): Add buffer_size.
(buf_read): Use s->buffer_size instead of BUFFER_SIZE.
(buf_write): Likewise.
(buf_init): Add argument unformatted.  Handle block sizes
for unformatted vs. formatted, using defaults if provided.
(fd_to_stream): Add argument unformatted in call to buf_init.
* libgfortran.h (options_t): Add buffer_size_formatted and
buffer_size_unformatted.
* runtime/environ.c (variable_table): Add
GFORTRAN_UNFORMATTED_BUFFER_SIZE and
GFORTRAN_FORMATTED_BUFFER_SIZE.

From-SVN: r273643

Daily bump.

From-SVN: r273640

rs6000: Make offsettable_mem_operand use any_memory_operand

* config/rs6000/predicates.md (offsettable_mem_operand): Allow volatile
memory.

From-SVN: r273633

rs6000: Make input_operand use any_memory_operand

* config/rs6000/predicates.md (input_operand): Allow volatile memory.

From-SVN: r273632

rs6000: Make lwa_operand use any_memory_operand

Testcase from comex, see https://lwn.net/Articles/793932/ .

* config/rs6000/predicates.md (lwa_operand): Allow volatile memory.

gcc/testsuite/
* gcc.target/powerpc/volatile-mem.c: New testcase.

From-SVN: r273631

rs6000: New predicate any_memory_operand

The new predicate accepts both memory_operand and volatile_mem_operand.

* config/rs6000/predicates.md (volatile_mem_operand): Modernize syntax.
(any_memory_operand): New predicate.
(reg_or_mem_operand): Use it.

From-SVN: r273630

re PR target/91204 (ICE in expand_expr_real_2, at expr.c:9215 with -O3)

PR target/91204
* optabs.c (expand_unop): As fallback, expand ~op0 as op0 ^ -1.

* gcc.c-torture/compile/pr91204.c: New test.

From-SVN: r273629

pa.h (hppa_profile_hook): Delete declaration.

* config/pa/pa.h (hppa_profile_hook): Delete declaration.
* config/pa/pa-protos.h (hppa_profile_hook): Add declaration.

From-SVN: r273628

Fix ICE on class template argument deduction with inherited ctor.

In general, when we see a dependent using-declaration we don't know whether
it names a function or not, so it doesn't get an OVERLOAD unless we see
overloads of the same name in the current class.  In the case of an
inherited constructor we could figure that out from the name, but it's
simpler to handle USING_DECL properly.

* cp-tree.h (ovl_iterator::using_p): A USING_DECL by itself was also
introduced by a using-declaration.

From-SVN: r273623

Reduce memory consumption for push/pop_access_scope.

I was seeing memory consumption issues on the concepts-cxx2a
branch. push_scope was, surprisingly, at the top of -fmem-report, and
push_access_scope was pretty high.  Fixing them was pretty simple.

* name-lookup.c (leave_scope): Do add class levels other than
previous_class_level to free_binding_level.
(invalidate_class_lookup_cache): Move from class.c, add to
free_binding_level.
* pt.c (saved_access_scope): Change from list to vec.

From-SVN: r273622

tree.def (OMP_LOOP): New tree code.

* tree.def (OMP_LOOP): New tree code.
* tree-core.h (enum omp_clause_code): Add OMP_CLAUSE_BIND.
(enum omp_clause_bind_kind): New enum.
(struct tree_omp_clause): Add subcode.bind_kind.
* tree.h (OMP_LOOP_CHECK): Rename to ...
(OMP_LOOPING_CHECK): ... this.
(OMP_FOR_BODY, OMP_FOR_CLAUSES, OMP_FOR_INIT, OMP_FOR_COND,
OMP_FOR_INCR, OMP_FOR_PRE_BODY, OMP_FOR_ORIG_DECLS): Use
OMP_LOOPING_CHECK instead of OMP_LOOP_CHECK.
(OMP_CLAUSE_BIND_KIND): Define.
* tree.c (omp_clause_num_ops, omp_clause_code_name): Add
bind clause entries.
(walk_tree_1): Handle OMP_CLAUSE_BIND.
* tree-pretty-print.c (dump_omp_clause): Likewise.
(dump_generic_node): Handle OMP_LOOP.
* gimplify.c (enum omp_region_type): Add ORT_IMPLICIT_TARGET.
(in_omp_construct): New variable.
(is_gimple_stmt): Handle OMP_LOOP.
(gimplify_scan_omp_clauses): For lastprivate don't set
check_non_private if code == OMP_LOOP.  For reduction clause
on OMP_LOOP combined with parallel or teams propagate as shared
on the combined construct.  Handle OMP_CLAUSE_BIND.
(gimplify_adjust_omp_clauses): Handle OMP_CLAUSE_BIND.
(gimplify_omp_for): Pass OMP_LOOP instead of OMP_{FOR,DISTRIBUTE}
for constructs from a loop construct to gimplify_scan_omp_clauses.
Don't predetermine iterator linear on OMP_SIMD from loop construct.
(replace_reduction_placeholders, gimplify_omp_loop): New functions.
(gimplify_omp_workshare): Use ORT_IMPLICIT_TARGET instead of trying
to match the implicit ORT_TARGET construct around whole body.
Temporarily clear in_omp_construct when processing body.
(gimplify_expr): Handle OMP_LOOP.  For OMP_MASTER, OMP_TASKGROUP
etc. temporarily set in_omp_construct when processing body.
(gimplify_body): Create ORT_IMPLICIT_TARGET instead of ORT_TARGET.
* omp-low.c (struct omp_context): Add loop_p.
(build_outer_var_ref): Treat ctx->loop_p similarly to simd construct
in that the original var might be private.
(scan_sharing_clauses): Handle OMP_CLAUSE_BIND.
(check_omp_nesting_restrictions): Adjust nesting restrictions for
addition of loop construct.
(scan_omp_1_stmt): Allow setjmp inside of loop construct.
gcc/c-family/
* c-pragma.h (enum pragma_kind): Add PRAGMA_OMP_LOOP.
(enum pragma_omp_clause): Add PRAGMA_OMP_CLAUSE_BIND.
* c-pragma.c (omp_pragmas_simd): Add PRAGMA_OMP_LOOP entry.
* c-common.h (enum c_omp_clause_split): Add C_OMP_CLAUSE_SPLIT_LOOP.
* c-omp.c (c_omp_split_clauses): Add support for 4 new combined
constructs with the loop construct.
gcc/c/
* c-parser.c (c_parser_omp_clause_name): Handle bind clause.
(c_parser_omp_clause_bind): New function.
(c_parser_omp_all_clauses): Handle PRAGMA_OMP_CLAUSE_BIND.
(OMP_LOOP_CLAUSE_MASK): Define.
(c_parser_omp_loop): New function.
(c_parser_omp_parallel, c_parser_omp_teams): Handle parsing of
loop combined with parallel or teams.
(c_parser_omp_construct): Handle PRAGMA_OMP_LOOP.
* c-typeck.c (c_finish_omp_clauses): Handle OMP_CLAUSE_BIND.
gcc/cp/
* cp-tree.h (OMP_FOR_GIMPLIFYING_P): Use OMP_LOOPING_CHECK
instead of OMP_LOOP_CHECK.
* parser.c (cp_parser_omp_clause_name): Handle bind clause.
(cp_parser_omp_clause_bind): New function.
(cp_parser_omp_all_clauses): Handle PRAGMA_OMP_CLAUSE_BIND.
(OMP_LOOP_CLAUSE_MASK): Define.
(cp_parser_omp_loop): New function.
(cp_parser_omp_parallel, cp_parser_omp_teams): Handle parsing of
loop combined with parallel or teams.
(cp_parser_omp_construct): Handle PRAGMA_OMP_LOOP.
(cp_parser_pragma): Likewise.
* pt.c (tsubst_expr): Handle OMP_LOOP.
* semantics.c (finish_omp_clauses): Handle OMP_CLAUSE_BIND.
gcc/testsuite/
* c-c++-common/gomp/cancel-1.c: Adjust expected diagnostic wording.
* c-c++-common/gomp/clauses-1.c (foo, baz, bar): Add order(concurrent)
clause where allowed.  Add combined constructs with loop with all
possible clauses.
(qux): New function.
* c-c++-common/gomp/loop-1.c: New test.
* c-c++-common/gomp/loop-2.c: New test.
* c-c++-common/gomp/loop-3.c: New test.
* c-c++-common/gomp/loop-4.c: New test.
* c-c++-common/gomp/loop-5.c: New test.
* c-c++-common/gomp/order-3.c: Adjust expected diagnostic wording.
* c-c++-common/gomp/simd-setjmp-1.c: New test.
* c-c++-common/gomp/teams-2.c: Adjust expected diagnostic wording.
libgomp/
* testsuite/libgomp.c-c++-common/loop-1.c: New test.

From-SVN: r273621

omp-low.c (lower_rec_input_clauses): Don't force simd arrays for lastprivate non-addressable iterator of a...

* omp-low.c (lower_rec_input_clauses): Don't force simd arrays for
lastprivate non-addressable iterator of a collapse(1) simd.

* gcc.dg/vect/vect-simd-16.c: New test.

From-SVN: r273620

Daily bump.

From-SVN: r273615

compiler: don't export bodies for functions marked "go:noinline"

    The current Mark_inline_candidates helper looks only at budget when
    deciding to mark a function or method as inline (with the proviso that
    IR constructs not yet supported by the inliner are given artificially
    high cost). This patch changes the helper to also look at whether a
    function has the "go:noinline" pragma; if it does have the pragma
    there is no point putting it into the export data (it will just make
    the export data bigger).

    Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/186923

From-SVN: r273611

rs6000-call.c (HAVE_AS_GNU_ATTRIBUTE): define value as in rs6000.c.

2019-07-17  Bill Seurer  <seurer@linux.vnet.ibm.com>

* config/rs6000/rs6000-call.c (HAVE_AS_GNU_ATTRIBUTE): define value
as in rs6000.c.

From-SVN: r273610

stl_tempbuf.h (__detail::__return_temporary_buffer): Fix sized deallocation size computation.

2019-07-19  François Dumont  <fdumont@gcc.gnu.org>

* include/bits/stl_tempbuf.h (__detail::__return_temporary_buffer): Fix
sized deallocation size computation.

From-SVN: r273609

[Darwin] More specs TLC.

This strips out a few driver specs that are only specifying a default state.
Also warn on an option now ignored, and add some comments to the driver specs
section.

2019-07-19  Iain Sandoe  <iain@sandoe.co.uk>

* config/darwin.h (DRIVER_SELF_SPECS): Ignore X and Mach specs which
refer to default conditions.  Warn for the 'y' spec which is ignored
by current linkers.

From-SVN: r273608

[PATCH, rs6000] Split up rs6000.c.

The source file rs6000.c has grown to unreasonable size and is being
split up into several smaller source files.  This should improve
compilation speed for building gcc.

This is the second of several patches to do this and moves most of the
function call and builtin code to a new source file.

Bootstrapped and tested on powerpc64le-unknown-linux-gnu and
powerpc64-unknown-linux-gnu with no regressions.  Is this ok for trunk?

2019-07-17  Bill Seurer  <seurer@linux.vnet.ibm.com>

* config/rs6000/rs6000.c (builtin_description, cpu_is_info,
cpu_supports_info, builtin_hash_struct, builtin_hasher,
builtin_hash_table, rs6000_builtin_info_type, rs6000_builtin_info,
rs6000_aggregate_candidate, rs6000_discover_homogeneous_aggregate,
rs6000_return_in_memory, rs6000_return_in_msb, call_ABI_of_interest,
init_cumulative_args, rs6000_promote_function_mode,
rs6000_must_pass_in_stack, is_complex_IBM_long_double,
abi_v4_pass_in_fpr, rs6000_function_arg_padding,
rs6000_function_arg_boundary, rs6000_parm_offset,
rs6000_parm_start, rs6000_arg_size,
rs6000_darwin64_record_arg_advance_flush,
rs6000_darwin64_record_arg_advance_recurse,
rs6000_darwin64_struct_check_p, rs6000_function_arg_advance_1,
rs6000_function_arg_advance, rs6000_darwin64_record_arg_flush,
rs6000_darwin64_record_arg_recurse, rs6000_darwin64_record_arg,
rs6000_mixed_function_arg, rs6000_psave_function_arg,
rs6000_finish_function_arg, rs6000_function_arg,
rs6000_arg_partial_bytes, rs6000_pass_by_reference,
rs6000_parm_needs_stack, rs6000_function_parms_need_stack,
rs6000_reg_parm_stack_space, rs6000_move_block_from_reg,
setup_incoming_varargs, rs6000_build_builtin_va_list, rs6000_va_start,
rs6000_gimplify_va_arg, def_builtin, bdesc_3arg, bdesc_dst,
bdesc_2arg, bdesc_altivec_preds, bdesc_abs, bdesc_1arg, bdesc_0arg,
bdesc_htm, rs6000_overloaded_builtin_p, rs6000_overloaded_builtin_name,
rs6000_expand_zeroop_builtin, rs6000_expand_mtfsf_builtin,
rs6000_expand_mtfsb_builtin, rs6000_expand_set_fpscr_rn_builtin,
rs6000_expand_set_fpscr_drn_builtin, rs6000_expand_unop_builtin,
altivec_expand_abs_builtin, rs6000_expand_binop_builtin,
altivec_expand_predicate_builtin, swap_endian_selector_for_mode,
altivec_expand_lv_builtin, altivec_expand_stxvl_builtin,
altivec_expand_stv_builtin, htm_spr_num, rs6000_htm_spr_icode,
htm_expand_builtin, cpu_expand_builtin, rs6000_expand_ternop_builtin,
altivec_expand_dst_builtin, altivec_expand_vec_init_builtin,
get_element_number, altivec_expand_vec_set_builtin,
altivec_expand_vec_ext_builtin, altivec_expand_builtin,
rs6000_builtin_is_supported_p, rs6000_invalid_builtin,
rs6000_fold_builtin, rs6000_builtin_valid_without_lhs,
fold_build_vec_cmp, fold_compare_helper, fold_mergehl_helper,
fold_mergeeo_helper, rs6000_gimple_fold_builtin,
rs6000_expand_builtin, rs6000_vector_type,
rs6000_init_builtins, rs6000_builtin_decl, altivec_init_builtins,
htm_init_builtins, builtin_function_type, rs6000_common_init_builtins,
rs6000_internal_arg_pointer, rs6000_output_mi_thunk): Move
to rs6000-call.c.
* config/rs6000/rs6000-call.c (builtin_description, cpu_is_info,
cpu_supports_info, builtin_hash_struct, builtin_hasher,
builtin_hash_table, rs6000_builtin_info_type, rs6000_builtin_info,
rs6000_aggregate_candidate, rs6000_discover_homogeneous_aggregate,
rs6000_return_in_memory, rs6000_return_in_msb, call_ABI_of_interest,
init_cumulative_args, rs6000_promote_function_mode,
rs6000_must_pass_in_stack, is_complex_IBM_long_double,
abi_v4_pass_in_fpr, rs6000_function_arg_padding,
rs6000_function_arg_boundary, rs6000_parm_offset,
rs6000_parm_start, rs6000_arg_size,
rs6000_darwin64_record_arg_advance_flush,
rs6000_darwin64_record_arg_advance_recurse,
rs6000_darwin64_struct_check_p, rs6000_function_arg_advance_1,
rs6000_function_arg_advance, rs6000_darwin64_record_arg_flush,
rs6000_darwin64_record_arg_recurse, rs6000_darwin64_record_arg,
rs6000_mixed_function_arg, rs6000_psave_function_arg,
rs6000_finish_function_arg, rs6000_function_arg,
rs6000_arg_partial_bytes, rs6000_pass_by_reference,
rs6000_parm_needs_stack, rs6000_function_parms_need_stack,
rs6000_reg_parm_stack_space, rs6000_move_block_from_reg,
setup_incoming_varargs, rs6000_build_builtin_va_list, rs6000_va_start,
rs6000_gimplify_va_arg, def_builtin, bdesc_3arg, bdesc_dst,
bdesc_2arg, bdesc_altivec_preds, bdesc_abs, bdesc_1arg, bdesc_0arg,
bdesc_htm, rs6000_overloaded_builtin_p, rs6000_overloaded_builtin_name,
rs6000_expand_zeroop_builtin, rs6000_expand_mtfsf_builtin,
rs6000_expand_mtfsb_builtin, rs6000_expand_set_fpscr_rn_builtin,
rs6000_expand_set_fpscr_drn_builtin, rs6000_expand_unop_builtin,
altivec_expand_abs_builtin, rs6000_expand_binop_builtin,
altivec_expand_predicate_builtin, swap_endian_selector_for_mode,
altivec_expand_lv_builtin, altivec_expand_stxvl_builtin,
altivec_expand_stv_builtin, htm_spr_num, rs6000_htm_spr_icode,
htm_expand_builtin, cpu_expand_builtin, rs6000_expand_ternop_builtin,
altivec_expand_dst_builtin, altivec_expand_vec_init_builtin,
get_element_number, altivec_expand_vec_set_builtin,
altivec_expand_vec_ext_builtin, altivec_expand_builtin,
rs6000_builtin_is_supported_p, rs6000_invalid_builtin,
rs6000_fold_builtin, rs6000_builtin_valid_without_lhs,
fold_build_vec_cmp, fold_compare_helper, fold_mergehl_helper,
fold_mergeeo_helper, rs6000_gimple_fold_builtin,
rs6000_expand_builtin, rs6000_vector_type,
rs6000_init_builtins, rs6000_builtin_decl, altivec_init_builtins,
htm_init_builtins, builtin_function_type, rs6000_common_init_builtins,
rs6000_internal_arg_pointer, rs6000_output_mi_thunk: Move
to here from rs6000.c.
* config/rs6000/rs6000-internal.h: (rs6000_darwin64_struct_check_p,
rs6000_discover_homogeneous_aggregate, rs6000_output_mi_thunk,
rs6000_output_addr_const_extra, rs6000_gimple_fold_builtin,
rs6000_invalid_builtin, rs6000_build_builtin_va_list, rs6000_va_start,
rs6000_gimplify_va_arg, rs6000_promote_function_mode,
rs6000_return_in_memory, rs6000_return_in_msb,
rs6000_pass_by_reference, setup_incoming_varargs,
rs6000_function_arg_boundary, rs6000_must_pass_in_stack,
rs6000_arg_partial_bytes, rs6000_function_arg_advance,
rs6000_function_arg_padding, rs6000_function_arg,
rs6000_darwin64_record_arg, rs6000_internal_arg_pointer,
rs6000_init_builtins, rs6000_builtin_decl, rs6000_expand_builtin,
rs6000_fold_builtin, rs6000_passes_ieee128, rs6000_passes_float,
rs6000_passes_long_double, rs6000_passes_vector,
rs6000_returns_struct, cpu_builtin_p, tree builtin_mode_to_type,
altivec_builtin_mask_for_load) Add declarations.
* config/rs6000/t-rs6000: Add new source file rs6000-call.c.
* config/config.gcc: Add new source file rs6000-call.c to garbage
collector and extra_objs.

From-SVN: r273607

tree-ssa-dse.c (initialize_ao_ref_for_dse): Handle strncpy.

* tree-ssa-dse.c (initialize_ao_ref_for_dse): Handle
strncpy.  Drop some trivial dead code.
(maybe_trim_memstar_call): Handle strncpy.

* gcc.dg/tree-ssa/ssa-dse-37.c: New test.
* gcc.dg/tree-ssa/ssa-dse-38.c: New test.

From-SVN: r273606

re PR tree-optimization/91211 (wrong code with __builtin_memset() and __builtin_memcpy() at -O1 and above)

2019-07-19  Richard Biener  <rguenther@suse.de>

PR tree-optimization/91211
* tree-ssa-sccvn.c (vn_walk_cb_data::push_partial_def): Fix
memset encoding size.

* gcc.dg/torture/pr91211.c: New testcase.

From-SVN: r273605

re PR target/91204 (ICE in expand_expr_real_2, at expr.c:9215 with -O3)

PR target/91204
* config/i386/mmx.md (one_cmpl<mode>2): New expander.

From-SVN: r273604

re PR ipa/91194 (A suspicious condition in recursive_inlining)

PR ipa/91194
* ipa-inline.c (recursive_inlining): Fix limits check.

From-SVN: r273603

re PR tree-optimization/91200 (ICE on valid code at -O1: verify_ssa failed)

2019-07-19  Richard Biener  <rguenther@suse.de>

PR tree-optimization/91200
* tree-ssa-phiopt.c (cond_store_replacement): Check we have
no PHI nodes in middle-bb.

* gcc.dg/torture/pr91200.c: New testcase.

From-SVN: r273602

[AArch64] Rename +bitperm to +sve2-bitperm

After some discussion, we've decided to rename the +bitperm feature
flag to +sve2-bitperm, so that it's consistent with the other SVE2
feature flags.  The associated macro was already
__ARM_FEATURE_SVE2_BITPERM, so only the feature flag itself
needs to change.

2019-07-19  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* doc/invoke.texi: Rename the AArch64 +bitperm extension flag
to +sve-bitperm.
* config/aarch64/aarch64-option-extensions.def: Likewise.

From-SVN: r273600

re PR middle-end/91190 (ICE on valid code: in hashtab_chk_error, at hash-table.c:137)

PR middle-end/91190
* function.c (insert_temp_slot_address): Store into the hash table
a copy of address to avoid RTL sharing issues.

* gcc.c-torture/compile/pr91190.c: New test.

From-SVN: r273599

* config/abi/post/m68k-linux-gnu/baseline_symbols.txt: Update.

From-SVN: r273598

re PR tree-optimization/91207 (Wrong code with -O3)

2019-07-19  Richard Biener  <rguenther@suse.de>

PR tree-optimization/91207
Revert
2019-07-17  Richard Biener  <rguenther@suse.de>

PR tree-optimization/91178
* tree-vect-stmts.c (get_group_load_store_type): For SLP
loads with a gap larger than the vector size always use
VMAT_STRIDED_SLP.
(vectorizable_load): For VMAT_STRIDED_SLP with a permutation
avoid loading vectors that are only contained in the gap
and thus are not needed.

* gcc.dg/torture/pr91207.c: New testcase.

From-SVN: r273593

PR c++/90101 - dependent class non-type parameter.

We shouldn't complain that a dependent type is incomplete.

* pt.c (invalid_nontype_parm_type_p): Check for dependent class type.

From-SVN: r273592

PR c++/90098 - partial specialization and class non-type parms.

A non-type template parameter of class type used in an expression has
const-qualified type; the pt.c hunks deal with this difference from the
unqualified type of the parameter declaration.  WAhen we use such a
parameter as an argument to another template, we don't want to confuse
things by copying it, we should pass it straight through.  And we might as
well skip copying other classes in constant evaluation context in a
template, too; we'll get the copy semantics at instantiation time.

PR c++/90099
PR c++/90101
* call.c (build_converted_constant_expr_internal): Don't copy.
* pt.c (process_partial_specialization): Allow VIEW_CONVERT_EXPR
around class non-type parameter.
(unify) [TEMPLATE_PARM_INDEX]: Ignore cv-quals.

From-SVN: r273591

Daily bump.

From-SVN: r273590

stl_tempbuf.h (__detail::__return_temporary_buffer): New.

2019-07-18  François Dumont  <fdumont@gcc.gnu.org>

* include/bits/stl_tempbuf.h (__detail::__return_temporary_buffer): New.
(~_Temporary_buffer()): Use latter.
(_Temporary_buffer(_FIterator, size_type)): Likewise.

From-SVN: r273586

i386.md (*addqi_2_slp): Remove.

* config/i386/i386.md (*addqi_2_slp): Remove.
(*<code>qi_2_slp): Ditto.

From-SVN: r273583

Rename function.

2019-07-18  Michael Meissner  <meissner@linux.ibm.com>

* config/rs6000/predicates.md (prefixed_mem_operand): Call
rs6000_prefixed_address_mode_p instead of rs6000_prefixed_address.
* config/rs6000/rs6000-protos.h (rs6000_prefixed_address_mode_p):
Rename function from rs6000_prefixed_address.
* config/rs6000/rs6000.c (rs6000_prefixed_address_mode_p): Rename
function from rs6000_prefixed_address.

From-SVN: r273580

Update PowerPC compiler for pc-relative support.

2019-07-18  Michael Meissner  <meissner@linux.ibm.com>

* config/rs6000/aix.h (TARGET_HAS_TOC): Rename TARGET_TOC to
TARGET_HAS_TOC.
(TARGET_TOC): Likewise.
(TARGET_NO_TOC): Delete here, define TARGET_NO_TOC_OR_PCREL in
rs6000.h.
* config/rs6000/darwin.h (TARGET_HAS_TOC): Rename TARGET_TOC to
TARGET_HAS_TOC.
(TARGET_TOC): Likewise.
(TARGET_NO_TOC): Delete here, define TARGET_NO_TOC_OR_PCREL in
rs6000.h.
* config/rs6000/linux64.h (TARGET_HAS_TOC): Rename TARGET_TOC to
TARGET_HAS_TOC.
(TARGET_TOC): Likewise.
* config/rs6000/rs6000.c (rs6000_option_override_internal): Add
check to require -mcmodel=medium for pc-relative addressing.
(create_TOC_reference): Add assertion for TARGET_TOC.
(rs6000_legitimize_address): Use TARGET_NO_TOC_OR_PCREL instead of
TARGET_NO_TOC.
(rs6000_emit_move): Likewise.
(TOC_alias_set): Rename TOC alias set static variable from 'set'
to 'TOC_alias_set'.
(get_TOC_alias_set): Likewise.
(output_toc): Use TARGET_NO_TOC_OR_PCREL instead of
TARGET_NO_TOC.
(rs6000_can_eliminate): Likewise.
* config/rs6000/rs6000.h (TARGET_TOC): Define in terms of
TARGET_HAS_TOC and not pc-relative.
(TARGET_NO_TOC_OR_PCREL): New macro to replace TARGET_NO_TOC.
* config/rs6000/sysv4.h (TARGET_HAS_TOC): Rename TARGET_TOC to
TARGET_HAS_TOC.
(TARGET_TOC): Likewise.
(TARGET_NO_TOC): Delete here, define TARGET_NO_TOC_OR_PCREL in
rs6000.h.

From-SVN: r273579

re PR target/91188 (strict_low_part operations do not work)

PR target/91188
* config/i386/i386.md (*addqi_1_slp): Use register_operand predicate
for operand 0.  Do not use (match_dup) to match operand 1 with
operand 0.  Add check in insn constraint that either input operand
matches operand 0.  Use SWI12 mode iterator to also handle
HImode operands.
(*and<mode>_1_slp): Ditto.
(*<code>qi_1_slp): Ditto.
(*sub<mode>_1_slp): Use register_operand predicate for operand 0.
Do not use (match_dup) to match operand 1 with operand 0.  Add
check in insn constraint that operand 1 matches operand 0.
Use SWI12 mode iterator to also handle HImode operands.
(*ashl<mode>3_1_slp): Ditto.
(*<shift_insn><mode>3_1_slp): Ditto.
(*<rotate_insn><mode>3_1_slp): Ditto.

testsuite/ChangeLog:

PR target/91188
* gcc.target/i386/pr91188-1a.c: New test.
* gcc.target/i386/pr91188-1b.c: Ditto.
* gcc.target/i386/pr91188-1c.c: Ditto.
* gcc.target/i386/pr91188-2a.c: Ditto.
* gcc.target/i386/pr91188-2b.c: Ditto.
* gcc.target/i386/pr91188-2c.c: Ditto.

From-SVN: r273578

compiler: fix bug in importing blocks from inline functions

    This patch fixes a buglet in the function body importer. Add hooks for
    keeping a stack of blocks corresponding to the block nesting in the
    imported function. This ensures that local variables and temps wind up
    correctly scoped and don't introduce collisions.

    New test case for this problem in CL 186717.

    Fixes golang/go#33158.

    Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/186757

From-SVN: r273577

Makefile.rtl, [...]: Introduce a "STANDALONE" mode where C runtime files do not have any dependency...

* Makefile.rtl, expect.c, env.c, aux-io.c, mkdir.c, initialize.c,
cstreams.c, raise.c, tracebak.c, adadecode.c, init.c, raise-gcc.c,
argv.c, adaint.c, adaint.h, ctrl_c.c, sysdep.c, rtinit.c, cio.c,
seh_init.c, exit.c, targext.c: Introduce a "STANDALONE" mode where C
runtime files do not have any dependency on GCC include files.
Remove unnecessary includes.
Remove remaining references to VMS in runtime C file.
* runtime.h: new File.

From-SVN: r273576

[patch2/2][arm]: remove builtin expand for sha1

This patch removes the builtin expand handling for sha1h/c/m/p and
replaces it with expand patterns. This should make it more consistent
with how we handle intrinsic implementations and cleans up the custom
sha1 code in the arm_expand builtins for unop and ternop.

2019-07-18  Sylvia Taylor  <sylvia.taylor@arm.com>

* config/arm/arm-builtins.c
(arm_expand_ternop_builtin): Remove explicit sha1 builtin handling.
(arm_expand_unop_builtin): Likewise.
* config/arm/crypto.md
(crypto_sha1h): Convert from define_insn to define_expand.
(crypto_<crypto_pattern>): Likewise.
(crypto_sha1h_lb): New define_insn.
(crypto_<crypto_pattern>_lb): Likewise.

From-SVN: r273575

[patch1/2][arm][PR90317]: fix sha1 patterns

This patch fixes:

1) Ice message thrown when using the crypto_sha1h intrinsic due to
incompatible mode used for zero_extend. Removed zero extend as it is
not a good choice for vector modes and using an equivalent single
mode like TI (128bits) instead of V4SI produces extra instructions
making it inefficient.

This affects gcc version 8 and above.

2) Incorrect combine optimizations made due to vec_select usage
in the sha1 patterns on arm. The patterns should only combine
a vec select within a sha1h<op> instruction when the lane is 0.

This affects gcc version 5 and above.

- Fixed by explicitly declaring the valid const int for such
optimizations. For cases when the lane is not 0, the vector
lane selection now occurs in a e.g. vmov instruction prior
to sha1h<op>.

- Updated the sha1h testcases on arm to check for additional
cases with custom vector lane selection.

The intrinsic functions for the sha1 patterns have also been
simplified which seems to eliminate extra vmovs like:
- vmov.i32 q8, #0.

2019-07-18  Sylvia Taylor  <sylvia.taylor@arm.com>

        PR target/90317
        * config/arm/arm_neon.h
        (vsha1h_u32): Refactor.
        (vsha1cq_u32): Likewise.
        (vsha1pq_u32): Likewise.
        (vsha1mq_u32): Likewise.
        * config/arm/crypto.md:
        (crypto_sha1h): Remove zero extend, correct vec select.
        (crypto_sha1c): Correct vec select.
        (crypto_sha1m): Likewise.
        (crypto_sha1p): Likewise.

        * gcc.target/arm/crypto-vsha1cq_u32.c (foo): Change return type to
        uint32_t.
        (GET_LANE, TEST_SHA1C_VEC_SELECT): New.
        * gcc.target/arm/crypto-vsha1h_u32.c (foo): Change return type to
        uint32_t.
        (GET_LANE, TEST_SHA1H_VEC_SELECT): New.
        * gcc.target/arm/crypto-vsha1mq_u32.c (foo): Change return type to
        uint32_t.
        (GET_LANE, TEST_SHA1M_VEC_SELECT): New.
        * gcc.target/arm/crypto-vsha1pq_u32.c (foo): Change return type to
        uint32_t.
        (GET_LANE, TEST_SHA1P_VEC_SELECT): New.

From-SVN: r273574

demangle.h (rust_is_mangled): Move to libiberty/rust-demangle.h.

include/
* demangle.h (rust_is_mangled): Move to libiberty/rust-demangle.h.
(rust_demangle_sym): Move to libiberty/rust-demangle.h.
libiberty/
* cplus-dem.c: Include rust-demangle.h.
* rust-demangle.c: Include rust-demangle.h.
* rust-demangle.h: New file.

From-SVN: r273573

[arm] Fix incorrect modes with 'borrow' operations

Looking through the arm backend I noticed that the modes used to pass
comparison types into subtract-with-carry operations were being
incorrectly set.  The result is that the compiler is not truly
self-consistent.  To clean this up I've introduced a new predicate,
arm_borrow_operation (borrowed from the AArch64 backend) which can
match the comparison type with the required mode and then fixed all
the patterns to use this.  The split patterns that were generating
incorrect modes have all obviously been fixed as well.

The basic rule for the use of a borrow is:
- if the condition code was set by a 'subtract-like' operation (subs, cmp),
  then use CCmode and LTU.
- if the condition code was by unsigned overflow of addition (adds), then
  use CC_Cmode and GEU.

* config/arm/predicates.md (arm_borrow_operation): New predicate.
* config/arm/arm.c (subdi3_compare1): Use CCmode for the split.
(arm_subdi3, subdi_di_zesidi, subdi_di_sesidi): Likewise.
(subdi_zesidi_zesidi): Likewise.
(negdi2_compare, negdi2_insn): Likewise.
(negdi_extensidi): Likewise.
(negdi_zero_extendsidi): Likewise.
(arm_cmpdi_insn): Likewise.
(subsi3_carryin): Use arm_borrow_operation.
(subsi3_carryin_const): Likewise.
(subsi3_carryin_const0): Likewise.
(subsi3_carryin_compare): Likewise.
(subsi3_carryin_compare_const): Likewise.
(subsi3_carryin_compare_const0): Likewise.
(subsi3_carryin_shift): Likewise.
(rsbsi3_carryin_shift): Likewise.
(negsi2_carryin_compare): Likewise.

From-SVN: r273572

lto-common.c (gimple_register_canonical_type_1): Do not look for non-ODR conflicts of types in anonymous namespaces.

* lto-common.c (gimple_register_canonical_type_1): Do not look for
non-ODR conflicts of types in anonymous namespaces.
(unify_scc): Do not merge anonymous namespace types.
* g++.dg/lto/alias-5_0.C: New testcase.
* g++.dg/lto/alias-5_1.C: New.
* g++.dg/lto/alias-5_2.c: New.

From-SVN: r273571

re PR tree-optimization/91137 (Wrong code with -O3)

        PR tree-optimization/91137
        * tree-ssa-loop-ivopts.c (struct ivopts_data): New field.
        (tree_ssa_iv_optimize_init, alloc_iv, tree_ssa_iv_optimize_finalize):
        Init, use and fini the above new field.
        (determine_base_object_1): New function.
        (determine_base_object): Reimplement using walk_tree.

gcc/testsuite
        PR tree-optimization/91137
        * gcc.c-torture/execute/pr91137.c: New test.

From-SVN: r273570

Make ifcvt clean up dead comparisons

This change is needed to avoid a regression in gcc.dg/ifcvt-3.c
for a later patch.  Without it, we enter CSE with a dead comparison left
by if-conversion and then eliminate the second (live) comparison in
favour of the dead one.  That's functionally correct in itself, but it
meant that we'd combine the subtraction and comparison into a SUBS
before we have a chance to fold away the subtraction.

2019-07-18  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* basic-block.h (CLEANUP_FORCE_FAST_DCE): New macro.
* cfgcleanup.c (cleanup_cfg): Call run_fast_dce if
CLEANUP_FORCE_FAST_DCE is set.
* ifcvt.c (rest_of_handle_if_conversion): Pass
CLEANUP_FORCE_FAST_DCE to the final cleanup_cfg call if
if-conversion succeeded.

From-SVN: r273569

Fix -Wreturn-type for static naked functions in C

This patch extends the fix for PR53633 to include static functions,
which were giving a bogus -Wreturn-type warning for C but not for C++.

2019-07-18  Richard Sandiford  <richard.sandiford@arm.com>

gcc/c/
PR c/53633
* c-decl.c (finish_function): Check targetm.warn_func_return
before issuing a -Wreturn-type warning.

gcc/testsuite/
* c-c++-common/pr53633-2.c: New test.

From-SVN: r273568

tree-ssa-sccvn.c (vn_walk_cb_data::push_partial_def): Refactor branches to make code less indented.

2019-07-18  Richard Biener  <rguenther@suse.de>

* tree-ssa-sccvn.c (vn_walk_cb_data::push_partial_def): Refactor
branches to make code less indented.

From-SVN: r273567

compiler: fix bug in handling of unordered set during exporting

    In CL 183850 a change was made to combine tracking/discovery of
    exported types and imported packages during export data generation. As
    a result of this refactoring a bug was introduced: the new code can
    potentially insert items into the exports set (an unordered_set) while
    iterating through the same set, which is illegal according to the spec
    for std::unordered_set.

    This patch fixes the problem by changing the type discovery phase to
    iterate through a separate list of sorted exports, as opposed to
    iterating through the main unordered set.  Also included is a change
    to fix the code that looks for variables that are referenced from
    inlined routine bodies (this code wasn't scanning all of the function
    that it needed to scan).

    New test case for this problem in CL 186697.

    Updates golang/go#33020.

    Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/185977

From-SVN: r273564

-Wmissing-attributes: check that we avoid duplicates and false positives

The initial patch for PR 81824 fixed various possibilities of
-Wmissing-attributes reporting duplicates and false positives.  The
test that avoided them was a little obscure, though, so this patch
rewrites it into a more self-evident form.

The patch also adds a testcase that already passed, but that
explicitly covers some of the possibilities of reporting duplicates
and false positives that preexisting tests did not cover.

for  gcc/ChangeLog

PR middle-end/81824
* attribs.c (decls_mismatched_attributes): Simplify the logic
that avoids duplicates and false positives.

for  gcc/testsuite/ChangeLog

PR middle-end/81824
* g++.dg/Wmissing-attributes-1.C: New.  Some of its fragments
are from Martin Sebor.

From-SVN: r273563

Daily bump.

From-SVN: r273562

pa.c (pa_som_asm_init_sections): Don't force all constant data into data section when generating PIC code.

* config/pa/pa.c (pa_som_asm_init_sections): Don't force all constant
data into data section when generating PIC code.
(pa_select_section): Use pa_reloc_rw_mask() to qualify relocs.
(pa_reloc_rw_mask): Return 3 when generating PIC code and when
generating code for SOM targets earlier than HP-UX 11.  Otherwise,
return 2 for SOM and 0 for other targets.

From-SVN: r273557

tree-ssa-dse.c (initialize_ao_ref_for_dse): Fix formatting.

        * tree-ssa-dse.c (initialize_ao_ref_for_dse): Fix formatting.
        (dse_walker::dse_optimize_stmt): Likewise.  Add missing return to
        avoid unexpected switch statement fallthru.

From-SVN: r273556

i386.md (*add<dwi>3_doubleword): Remove redundant constraints.

* config/i386/i386.md (*add<dwi>3_doubleword):
Remove redundant constraints.
(*add<mode>_1): Ditto.
(*addhi_1): Ditto.
(*addqi_1): Ditto.
(*addqi_1_slp): Ditto.
(*add<mode>_2): Ditto.
(*addv<mode>4): Ditto.
(*sub<dwi>3_doubleword): Ditto.
(*sub<mode>_1): Ditto.
(*subqi_1_slp): Ditto.
(*sub<mode>_2): Ditto.
(*subv<mode>4): Ditto.
(*sub<mode>_3): Ditto.
(@add<mode>3_carry): Ditto.
(@sub<mode>3_carry): Ditto.
(*add<mode>3_cc_overflow_1): Ditto.
(*add<mode>3_zext_cc_overflow_2): Ditto.
(*anddi_1): Ditto.
(*and<mode>_1): Ditto.
(*andqi_1): Ditto.
(*andqi_1_slp): Ditto.
(*anddi_2): Ditto.
(*andqi_2_maybe_si): Ditto.
(*and<mode>_2): Ditto.
(*andqi_2_slp): Ditto.
(*<code><mode>_1): Ditto.
(*<code>qi_1): Ditto.
(*<code>qi_1_slp): Ditto.
(*<code><mode>_2): Ditto.
(*<code>qi_2_slp): Ditto.

From-SVN: r273554

re PR c++/90455 (braced-init and incomplete type instantiation)

PR c++/90455
* g++.dg/cpp0x/nsdmi-list6.C: New test.

From-SVN: r273553

alias.c (record_component_aliases): Do not simplify pointed-to types of ODR types

* alias.c (record_component_aliases): Do not simplify pointed-to
types of ODR types
* testsuite/g++.dg/lto/alias-4_0.C

From-SVN: r273552

i386.md (*andqi_2_maybe_si): Handle potential partial reg stall on alternative 2.

* config/i386/i386.md (*andqi_2_maybe_si): Handle potential
partial reg stall on alternative 2.

From-SVN: r273551

re PR tree-optimization/91178 (Infinite recursion in split_constant_offset in slp after r260289)

2019-07-17  Richard Biener  <rguenther@suse.de>

PR tree-optimization/91178
* tree-ssa.c (release_defs_bitset): Iterate from higher to
lower SSA names to avoid quadratic behavior in the common case.
* tree-data-ref.c (split_constant_offset): Add limit argument
and pass it down.  Initialize it from PARAM_SSA_NAME_DEF_CHAIN_LIMIT.
(split_constant_offset_1): Add limit argument and use it to
limit SSA def walking.  Optimize the common plus/minus case.

From-SVN: r273550

re PR tree-optimization/91178 (Infinite recursion in split_constant_offset in slp after r260289)

2019-07-17  Richard Biener  <rguenther@suse.de>

PR tree-optimization/91178
* tree-vect-stmts.c (get_group_load_store_type): For SLP
loads with a gap larger than the vector size always use
VMAT_STRIDED_SLP.
(vectorizable_load): For VMAT_STRIDED_SLP with a permutation
avoid loading vectors that are only contained in the gap
and thus are not needed.

* gcc.dg/torture/pr91178.c: New testcase.

From-SVN: r273549

re PR tree-optimization/91180 (wrong code at -O and above with __builtin_memset())

2019-07-17  Richard Biener  <rguenther@suse.de>

PR tree-optimization/91180
* tree-ssa-sccvn.c (vn_reference_lookup_3): Fix offset
computation for memset partial defs.

* gcc.dg/torture/pr91180.c: New testcase.

From-SVN: r273548

* config/abi/post/ia64-linux-gnu/baseline_symbols.txt: Update.

From-SVN: r273547

re PR tree-optimization/91157 (ICE: verify_gimple failed (error: position plus size exceeds size of referenced object in 'bit_field_ref'))

PR tree-optimization/91157
* tree-vect-generic.c (expand_vector_comparison): Handle lhs being
a vector boolean with scalar mode.
(expand_vector_condition): Handle first operand being a vector boolean
with scalar mode.
(expand_vector_operations_1): For comparisons, don't bail out early
if the return type is vector boolean with scalar mode, but comparison
operand type is not.

* gcc.target/i386/avx512f-pr91157.c: New test.
* gcc.target/i386/avx512bw-pr91157.c: New test.

From-SVN: r273545

gimple.h (enum gf_mask): Remove GF_OMP_FOR_SIMD...

* gimple.h (enum gf_mask): Remove GF_OMP_FOR_SIMD, change
GF_OMP_FOR_KIND_SIMD to a value serially after other kinds,
divide GF_OMP_FOR_KIND_MASK, GF_OMP_FOR_COMBINED,
GF_OMP_FOR_COMBINED_INTO, GF_OMP_FOR_GRID_PHONY,
GF_OMP_FOR_GRID_INTRA_GROUP and GF_OMP_FOR_GRID_GROUP_ITER by two.
* omp-grid.c (grid_process_grid_body,
grid_eliminate_combined_simd_part): Use GF_OMP_FOR_KIND_SIMD instead
of GF_OMP_FOR_SIMD, don't test & GF_OMP_FOR_SIMD but
== GF_OMP_FOR_KIND_SIMD.
* omp-low.c (build_outer_var_ref, scan_sharing_clauses,
check_omp_nesting_restrictions, scan_omp_1_stmt,
lower_rec_input_clauses, lower_lastprivate_conditional_clauses,
lower_lastprivate_clauses, lower_reduction_clauses, lower_omp_scan,
omp_find_scan): Likewise.
* omp-expand.c (expand_omp_for): Likewise.
* omp-general.c (omp_extract_for_data): Likewise.

From-SVN: r273544

re PR tree-optimization/91157 (ICE: verify_gimple failed (error: position plus size exceeds size of referenced object in 'bit_field_ref'))

PR tree-optimization/91157
* tree-vect-generic.c (expand_vector_comparison): Handle lhs being
a vector boolean with scalar mode.
(expand_vector_condition): Handle first operand being a vector boolean
with scalar mode.
(expand_vector_operations_1): For comparisons, don't bail out early
if the return type is vector boolean with scalar mode, but comparison
operand type is not.

* gcc.target/i386/avx512f-pr91157.c: New test.
* gcc.target/i386/avx512bw-pr91157.c: New test.

From-SVN: r273543

re PR tree-optimization/91181 (Failing as_as type conversion in vect_build_slp_tree_1)

2019-07-17  Richard Biener  <rguenther@suse.de>

PR tree-optimization/91181
* tree-vect-slp.c (vect_build_slp_tree_1): Do not compare
IFN_LOADs as calls.

* gcc.dg/pr91181.c: New testcase.

From-SVN: r273542

Daily bump.

From-SVN: r273541

re PR fortran/90903 (Implement runtime checks for bit manipulation intrinsics)

2019-07-16  Harald Anlauf  <anlauf@gmx.de>

PR fortran/90903
* libgfortran.h: Add mask for -fcheck=bits option.
* options.c (gfc_handle_runtime_check_option): Add option "bits"
to run-time checks selectable via -fcheck.
* trans-intrinsic.c (gfc_conv_intrinsic_btest)
(gfc_conv_intrinsic_singlebitop, gfc_conv_intrinsic_ibits)
(gfc_conv_intrinsic_shift, gfc_conv_intrinsic_ishft)
(gfc_conv_intrinsic_ishftc): Implement run-time checks for the
POS, LEN, SHIFT, and SIZE arguments.
* gfortran.texi: Document run-time checks for bit manipulation
intrinsics.
* invoke.texi: Document new -fcheck=bits option.

PR fortran/90903
* gfortran.dg/check_bits_1.f90: New testcase.

From-SVN: r273535

i386.md (*testdi_1): Match CCZmode for constants that might have the SImode sign bit set.

* config/i386/i386.md (*testdi_1): Match CCZmode for
constants that might have the SImode sign bit set.
(*testqi_1_maybe_si): Remove "!" constraint modifier.
Use correct constraints for pentium pairing.
(*test<mode>_1): Ditto.

From-SVN: r273534

re PR rtl-optimization/91173 (ICE: in int_mode_for_mode, at stor-layout.c:403)

PR rtl-optimization/91173
* g++.dg/pr91173.C: New test.

From-SVN: r273531

re PR rtl-optimization/91173 (ICE: in int_mode_for_mode, at stor-layout.c:403)

PR rtl-optimization/91173
* tree-ssa-address.c (addr_for_mem_ref): If the base is an
SSA_NAME with a constant value, fold its value into the offset
and clear the base before calling gen_addr_rtx.

From-SVN: r273529

Add missing testcase for PR89190

    testsuite/
PR target/89190
* gcc.target/arm/pr89190.c: New test.

From-SVN: r273524

re PR rtl-optimization/91164 (ICE in verify_dominators, at dominance.c:1184 (error: dominator of 114 should be 112, not 16))

PR rtl-optimization/91164
* dse.c (rest_of_handle_dse): If dead edges have been purged,
invalidate dominance info.

* g++.dg/opt/pr91164.C: New test.

From-SVN: r273522

alias-1_0.C: Use -O3.

* g++.dg/lto/alias-1_0.C: Use -O3.
* g++.dg/lto/alias-2_0.C: Use -O3.
* g++.dg/lto/alias-3_0.C: Add loop to enable inlining with
-fno-use-linker-plugin.
* g++.dg/lto/alias-3_1.C: Remove dg-lto-do and dg-lto-options.

From-SVN: r273520

Don't xfail gcc.dg/tree-ssa/pr84512.c on 64-bit SPARC

* gcc.dg/tree-ssa/pr84512.c: Don't xfail scan-tree-dump on
sparcv9.

From-SVN: r273519

Require int128 support in gcc.dg/autopar/pr91162.c

* gcc.dg/autopar/pr91162.c: Require int128 support.

From-SVN: r273518

Simplify range location creation in C++ parser.

Many places in the parser follow the same pattern of capturing the location
of the last lexed token, either before or after lexing it, and then using
that as the end of a location range; this can be simplified by passing the
lexer to make_location and grabbing the token location there.

* parser.c (make_location): Add overload taking cp_lexer* as last
parameter.

From-SVN: r273516

Add parens around constraint in <memory>.

C++20 concepts require parens around atomic constraints that are not
primary-expressions.

* include/std/memory (uses_allocator_construction_args): Add parens
around constraint.

From-SVN: r273515

Simplify type-specifier parsing.

Previously, the tentative parses for optional type-specifier and to support
class template argument deduction were combined awkwardly.  This
reorganization was motivated by the new concepts branch.

* parser.c (cp_parser_simple_type_specifier): Separate tentative
parses for optional type-spec and CTAD.

From-SVN: r273514

Fix g++.dg/template/pr84789.C on new concepts branch.

On the concepts branch I ran into trouble where a pre-parsed dependent
nested-name-specifier got replaced on a subsequent parse with is_declaration
by one with typenames resolved, which was then used wrongly on a further
parse with !is_declaration.

    * parser.c (cp_parser_nested_name_specifier_opt): If the token is
    already CPP_NESTED_NAME_SPECIFIER, leave it alone.

From-SVN: r273513

Report ambiguous uses of .md attributes

This patch reports an error if the .md file has an unscoped
attribute that maps to more than one possible value.

2019-07-16  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* read-md.h (md_reader::record_potential_iterator_use): Add a
file_location parameter.
* read-rtl.c (attribute_use::loc): New field.
(map_attr_string): Take a file_location parameter.  Report cases
in which attributes map to multiple distinct values.
(apply_attribute_uses): Update call accordingly.
(md_reader::handle_overloaded_name): Likewise.
(md_reader::apply_iterator_to_string): Likewise.  Skip empty
nonnull strings.
(record_attribute_use): Take a file_location parameter.
Initialize attribute_use::loc.
(md_reader::record_potential_iterator_use): Take a file_location
parameter.  Update call to record_attribute_use.
(rtx_reader::rtx_alloc_for_name): Update call accordingly.
(rtx_reader::read_rtx_code): Likewise.
(rtx_reader::read_rtx_operand): Likewise.  Record a location
for implicitly-expanded empty strings.

From-SVN: r273511

Use file_location for md_reader's ptr_loc

Also make it public, so that clients can use the location for error
reporting.

2019-07-16  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* read-md.h (md_reader::ptr_loc): Moved from read-md.c.
Use file_location instead of separate fields.
(md_reader::set_md_ptr_loc): Take a file_location instead of a
separate filename and line number.
* read-md.c (ptr_loc): As above.
(md_reader::copy_md_ptr_loc): Update for new ptr_loc layout.
(md_reader::fprint_md_ptr_loc): Likewise.
(md_reader::set_md_ptr_loc): Likewise.  Take a file_location
instead of a separate filename and line number.
(md_reader::read_string): Update call accordingly.

From-SVN: r273510

[rs6000] Fix ambiguous .md attribute uses

This patch is part of a series that fixes ambiguous attribute
uses in .md files, i.e. cases in which attributes didn't use
<ITER:ATTR> to specify an iterator, and in which <ATTR> could
have different values depending on the iterator chosen.

No behavioural change -- produces the same code as before except
for formatting and line numbers.

2019-07-16  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* config/rs6000/rs6000.md (*mov<mode>_update1): Explicitly
use <SFDF:mode>, <SFDF:MODE>, <SFDF:Ff> and <SFDF:bits> rather than
leaving the choice between SFDF and P implicit.
(*mov<mode>_update2): Likewise.
(*cmp<IBM128:mode>_internal2): Explicitly use <IBM128:MODE>
rather than leaving the choice betweem IBM128 and GPR implicit.
(*fix<uns>_trunc<IEEE128:mode><QHSI:mode>2_mem): Explicitly use
<IEEE128:MODE> rather than leaving the choice between IEEE128 and
QHSI implicit.
(AltiVec define_peephole2s): Explicitly use <ALTIVEC_DFORM:MODE>
rather than leaving the choice between ALTIVEC_DFORM and P implicit.
* config/rs6000/vsx.md
(*vsx_ext_<VSX_EXTRACT_I:VS_scalar>_fl_<FL_CONV:mode>)
(*vsx_ext_<VSX_EXTRACT_I:VS_scalar>_ufl_<FL_CONV:mode>): Explicitly
use <FL_CONV:VSisa> rather than leaving the choice between FL_CONV
and VSX_EXTRACT_I implicit.

From-SVN: r273509

[mips] Fix ambiguous .md attribute uses

This patch is part of a series that fixes ambiguous attribute
uses in .md files, i.e. cases in which attributes didn't use
<ITER:ATTR> to specify an iterator, and in which <ATTR> could
have different values depending on the iterator chosen.

No behavioural change -- produces the same code as before.

2019-07-16  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* config/mips/micromips.md (*movep<MOVEP1:mode><MOVEP2:mode>):
Explicitly use <MOVEP1:MODE> for the mode attribute.

From-SVN: r273508

PR91176: Skip debug insns when computing inline costs

Apply Honza's fix for an aarch64-linux-gnu bootstrap failure.

2019-07-16  Jan Hubicka  <hubicka@ucw.cz>

gcc/
PR bootstrap/91176
* ipa-fnsummary.c (analyze_function_body): Skip debug stmts

From-SVN: r273507

Remove trailing space.

From-SVN: r273506

Daily bump.

From-SVN: r273505

* config/i386/i386.md (*testdi_1): Put back "%" constraint modifier.

From-SVN: r273501

runtime: expose the g variable

    Currently, getg is implemented in C, which loads the thread-local
    g variable. The g variable is declared static in C.

    This CL exposes the g variable, so it can be accessed from the Go
    side. This allows the Go compiler to inline getg calls to direct
    access of g.

    Currently, the actual inlining is only implemented in the gollvm
    compiler. The g variable is thread-local and the compiler backend
    may choose to cache the TLS address in a register or on stack. If
    a thread switch happens the cache may become invalid. I don't
    know how to disable the TLS address cache in gccgo, therefore
    the inlining of getg is not implemented. In the future gccgo may
    gain this if we know how to do it safely.

    Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/186238

From-SVN: r273499

rs6000: Always output .machine

We now can always output .machine (if we output it at all for the
current target).

PR target/91050
* config/rs6000/rs6000.c (rs6000_file_start): Never skip emitting a
.machine directive.

From-SVN: r273498

i386.md (@test<mode>_ccno_1): Rename from test<mode>_ccno_1.

* config/i386/i386.md (@test<mode>_ccno_1):
Rename from test<mode>_ccno_1.
(*testdi_1): Remove redundant alternatives.  Remove modrm attribute.
(*testqi_1_maybe_si): Remove modrm attribute.
(*test<mode>_1): Ditto.
* config/i386/i386-expand.c (ix86_split_idivmod): Use
gen_test_ccno_1 and gen_extend_insn.

From-SVN: r273496